Method for judging coating quality of graphene-taikong coating of water turbine

Abstract

The application discloses a kind of water turbine graphene ceramic gold coating coating spraying quality judging method, belong to image processing technical field.The application first extracts the darkness and brightness of each pixel point in water turbine graphene ceramic gold coating image, calculates the darkness difference value and brightness difference value under different window size;Based on the multi-scale difference value, mark the darkness abnormal point and brightness abnormal point, filter the maximum value of difference value corresponding to abnormal point to generate darkness abnormal distribution graph and brightness abnormal distribution graph;The above abnormal distribution graph is calibrated by region connectivity analysis, and the darkness abnormal calibration graph and brightness abnormal calibration graph are obtained;According to the difference value of two kinds of calibration graphs, the light-dark consistency index is calculated and the corresponding distribution graph is generated;Finally, the guided enhancement quality evaluation neural network is used to process the calibration graph, the light-dark consistency index distribution graph is combined for feature enhancement, and the coating spraying quality score is output.The application improves the accuracy of coating detection.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to a method for judging the coating quality of graphene ceramic gold coating on a water turbine. Background Technology

[0002] Existing automated coating quality inspection methods often employ image recognition techniques based on fixed-scale local features. The system first acquires a visible light image of the coating surface, performs preprocessing such as denoising and normalization, and then calculates the average pixel brightness, grayscale difference, or local texture descriptor within a pre-defined fixed-size window. Based on these calculations, a threshold or classifier is used to determine whether the central pixel of the window belongs to an abnormal region, thereby locating dark spots, bright spots, or uneven coating areas. Because this type of method only performs local discrimination at a single window scale, its feature representation and judgment criteria are highly dependent on the spatial scale of the defect. Therefore, fixed-scale local image features cannot simultaneously guarantee comprehensive detection when dealing with coating defects of different sizes, leading to insufficient detection accuracy. Summary of the Invention

[0003] To address the aforementioned shortcomings in the existing technology, the present invention provides a method for judging the coating quality of graphene ceramic gold coating for water turbines, which solves the problem of insufficient detection accuracy in the existing technology.

[0004] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: a method for judging the coating quality of graphene ceramic gold coating on a water turbine, comprising the following steps:

[0005] Darkness and brightness were extracted from each pixel in the image of the graphene ceramic coating of the water turbine, and the difference values ​​of darkness and brightness under different window sizes were obtained.

[0006] Based on multiple darkness difference values ​​and multiple lightness difference values ​​of the same pixel, darkness anomalies and lightness anomalies are marked, and the maximum value of multiple darkness difference values ​​and lightness difference values ​​of the anomalies is selected to obtain darkness anomaly distribution map and lightness anomaly distribution map.

[0007] Based on the regional connectivity of the dark anomaly points and the light anomaly points, the dark anomaly distribution map and the light anomaly distribution map are calibrated respectively to obtain the dark anomaly calibration map and the light anomaly calibration map.

[0008] Based on the difference between the darkness anomaly calibration map and the lightness anomaly calibration map, the lightness uniformity index is generated, and the lightness uniformity index distribution map is obtained.

[0009] A guided enhanced quality assessment neural network was used to process the dark anomaly calibration map and the light anomaly calibration map, and feature enhancement was performed based on the light and dark consistency index distribution map to obtain the coating quality score.

[0010] Furthermore, the process of obtaining the darkness difference value includes:

[0011] For each pixel's R, G, and B channel values, select the minimum channel value and use the ratio of the minimum channel value to 255 as the darkness of the corresponding pixel.

[0012] Centered on each pixel, calculate the mean and standard deviation of darkness for each type of window to obtain the mean and standard deviation of window darkness;

[0013] Based on the difference between the darkness of each pixel and the mean darkness of the window, and normalized using the standard deviation of window darkness, the darkness difference value under the corresponding window is obtained.

[0014] Furthermore, the formula for obtaining the difference in darkness is:

[0015] ,

[0016] in, For the first The first pixel The difference in darkness between the windows. It is a natural constant. For the first The darkness of each pixel For the first The first pixel The average window darkness of each type of window. For the first The first pixel The standard deviation of window darkness for this type of window. For the denominator parameter, and `|` represents a positive integer, and `|` represents the absolute value operation.

[0017] Furthermore, the process of obtaining the brightness difference value includes:

[0018] For each pixel's R, G, and B channel values, select the maximum channel value and use the ratio of the maximum channel value to 255 as the brightness of the corresponding pixel.

[0019] Centered on each pixel, calculate the mean and standard deviation of brightness for each type of window to obtain the mean and standard deviation of window brightness.

[0020] The brightness difference value is obtained by normalizing the difference between the brightness of each pixel and the mean brightness of the window using the standard deviation of the window brightness.

[0021] Furthermore, the formula for obtaining the brightness difference value is:

[0022] ,

[0023] in, For the first The first pixel The difference in brightness between the windows. It is a natural constant. For the first The brightness of each pixel For the first The first pixel The average brightness of the window type. For the first The first pixel The standard deviation of window brightness for each type of window. For the denominator parameter, and `|` represents a positive integer, and `|` represents the absolute value operation.

[0024] Furthermore, the process of obtaining the darkness anomaly distribution map and the lightness anomaly distribution map includes:

[0025] For any pixel, if there exists a window whose darkness difference value is greater than a threshold, then the pixel is marked as a darkness outlier.

[0026] For any pixel, if there exists a window whose brightness difference value is greater than a threshold, then the pixel is marked as a brightness outlier.

[0027] The maximum value is selected from the darkness difference values ​​of multiple windows corresponding to the darkness anomaly points and used as the representative value of the darkness difference of the corresponding points. The other normal pixels are set to 0 to obtain the darkness anomaly distribution map.

[0028] The maximum value is selected from the brightness difference values ​​of multiple windows corresponding to the brightness anomaly points and used as the representative value of the brightness difference of the corresponding points. The other normal pixels are set to 0 to obtain the brightness anomaly distribution map.

[0029] Furthermore, the process of obtaining the darkness anomaly calibration map and the lightness anomaly calibration map includes:

[0030] The number of dark anomalies within a 3×3 neighborhood of a dark anomaly is counted, and the ratio of the number of dark anomalies to the number of pixels in the 3×3 neighborhood is taken as the region connectivity of that dark anomaly.

[0031] The darkness difference representative value of the same darkness anomaly point in the darkness anomaly distribution map is calibrated based on the regional connectivity of the darkness anomaly points to obtain the darkness anomaly calibration map.

[0032] The number of brightness anomalies within a 3×3 neighborhood of a brightness anomaly is counted, and the ratio of the number of brightness anomalies to the number of pixels in the 3×3 neighborhood is taken as the region connectivity of that brightness anomaly.

[0033] The brightness difference representative value of the same brightness anomaly point in the brightness anomaly distribution map is calibrated based on the regional connectivity of the brightness anomaly points to obtain the brightness anomaly calibration map.

[0034] Furthermore, the process of obtaining the uniformity of light and dark index distribution map includes:

[0035] At the same pixel location, if the corresponding pixel location in the darkness anomaly calibration map is a darkness anomaly point, or if the corresponding pixel location in the lightness anomaly calibration map is a lightness anomaly point, calculate the lightness / darkness consistency index: ,in, For the first The brightness uniformity index of each pixel For the first Calibration value for darkness difference of each pixel. For the first The brightness difference calibration value of each pixel, where e is the natural constant and || is the absolute value;

[0036] At the same pixel, the corresponding pixel position in the darkness anomaly calibration map is a normal pixel position, and the corresponding pixel position in the brightness anomaly calibration map is also a normal pixel position. Set the brightness uniformity index of the corresponding pixel position to 0 to obtain the brightness uniformity index distribution map.

[0037] Furthermore, the guided enhancement quality assessment neural network includes: a first guided enhancement module, a second guided enhancement module, a first CNN network, a second CNN network, an adder, and a fully connected layer.

[0038] Furthermore, the expression for the first boot enhancement module is:

[0039] ,

[0040] in, This is the darkness consistency enhancement feature output by the first guidance enhancement module. This is a calibration map for darkness anomalies. This is a distribution chart of the uniformity of light and dark indices. For element-wise multiplication, This is the first convolutional layer. This is the second convolutional layer;

[0041] The expression for the second boot enhancement module is:

[0042] ,

[0043] in, The brightness consistency enhancement feature output by the second guidance enhancement module. This is a brightness anomaly calibration chart. This is the third convolutional layer.

[0044] The beneficial effects of this invention are as follows:

[0045] 1. Existing methods rely on only a single window scale, resulting in poor adaptability to defects of different sizes. This invention extracts the dark / brightness difference value through multiple window scales while retaining the feature information of defects of different sizes. It can accurately identify both small-scale local defects and large-scale regional inhomogeneities, making defect detection more comprehensive and improving accuracy.

[0046] 2. This invention calibrates anomaly maps by using regional connectivity, which can effectively distinguish between "real continuous defects" and "isolated noise".

[0047] 3. This invention introduces a brightness uniformity index to quantify the spatial matching degree between darkness anomalies and brightness anomalies, directly reflecting the quality problem of "uniform brightness" in the coating; combined with the guided enhancement neural network for in-depth feature mining, it highlights the coupling relationship of the same location in different brightness directions, thereby significantly improving the discrimination significance of abnormal areas, suppressing background interference, and improving the accuracy of detection. Attached Figure Description

[0048] Figure 1 A flowchart illustrating a method for judging the quality of graphene ceramic coating spraying on a water turbine;

[0049] Figure 2 A schematic diagram of the structure of a neural network for guiding enhanced quality assessment. Detailed Implementation

[0050] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0051] like Figure 1 As shown, a method for judging the coating quality of graphene ceramic gold coating on a water turbine includes the following steps:

[0052] Darkness and brightness were extracted from each pixel in the image of the graphene ceramic coating of the water turbine, and the difference values ​​of darkness and brightness under different window sizes were obtained.

[0053] Based on multiple darkness difference values ​​and multiple lightness difference values ​​of the same pixel, darkness anomalies and lightness anomalies are marked, and the maximum value of multiple darkness difference values ​​and lightness difference values ​​of the anomalies is selected to obtain darkness anomaly distribution map and lightness anomaly distribution map.

[0054] Based on the regional connectivity of the dark anomaly points and the light anomaly points, the dark anomaly distribution map and the light anomaly distribution map are calibrated respectively to obtain the dark anomaly calibration map and the light anomaly calibration map.

[0055] Based on the difference between the darkness anomaly calibration map and the lightness anomaly calibration map, the lightness uniformity index is generated, and the lightness uniformity index distribution map is obtained.

[0056] A guided enhanced quality assessment neural network was used to process the dark anomaly calibration map and the light anomaly calibration map, and feature enhancement was performed based on the light and dark consistency index distribution map to obtain the coating quality score.

[0057] In this embodiment, the different window sizes include: 5×5 window, 7×7 window, 9×9 window and 11×11 window.

[0058] In this embodiment, the process of obtaining the darkness difference value includes:

[0059] For each pixel's R, G, and B channel values, select the minimum channel value and use the ratio of the minimum channel value to 255 as the darkness of the corresponding pixel.

[0060] ,

[0061] in, For the first The darkness of each pixel For the first The R channel value of each pixel For the first The G channel value of each pixel For the first The B channel value of each pixel It is a positive integer. To obtain the minimum value;

[0062] Centered on each pixel, a square neighborhood of size m×m is selected as a window. The mean and standard deviation of darkness under each window are calculated to obtain the mean and standard deviation of window darkness. The value of m is 5, 7, 9, and 11.

[0063] Based on the difference between the darkness of each pixel and the mean darkness of the window, and normalized using the standard deviation of window darkness, the darkness difference value under the corresponding window is obtained.

[0064] In this embodiment, the formula for obtaining the darkness difference value is:

[0065] ,

[0066] in, For the first The first pixel The difference in darkness between the windows. It is a natural constant. For the first The darkness of each pixel For the first The first pixel The average window darkness of each type of window. For the first The first pixel The standard deviation of window darkness for this type of window. For the denominator parameter, and `|` represents a positive integer, and `|` represents the absolute value operation.

[0067] This invention obtains the basic brightness and darkness features of a pixel by minimizing and normalizing the R, G, and B channels of each pixel. Furthermore, it calculates the local mean and standard deviation of darkness using a multi-scale m×m neighborhood window, and generates a darkness difference value by normalizing the difference between the pixel darkness and the window mean. This darkness difference value quantifies the degree of darkness aberration of a pixel relative to its neighborhood, making aberrant areas have a higher response in the numerical space, while normal areas have a lower response because their brightness is close to the neighborhood average.

[0068] This invention, by "calculating the mean / standard deviation of darkness under multiple windows", can simultaneously capture the differences between small windows (corresponding to tiny pinholes and localized missed spraying) and large windows (corresponding to large-area uneven spraying), solving the problem that traditional fixed-scale methods "can only adapt to defects of a single size", and achieving full coverage detection of defects of different sizes.

[0069] In this invention, when the difference between the darkness of a pixel and the window mean is small (normal area), the darkness difference value is close to 0, weakening irrelevant fluctuations. When the difference is large (true dark defects), the difference value approaches 1, highlighting defect features and improving the distinction between defects and normal areas.

[0070] In this embodiment, the process of obtaining the brightness difference value includes:

[0071] For each pixel's R, G, and B channel values, select the maximum channel value and use the ratio of the maximum channel value to 255 as the brightness of the corresponding pixel.

[0072] ,

[0073] in, For the first The brightness of each pixel For the first The R channel value of each pixel For the first The G channel value of each pixel For the first The B channel value of each pixel It is a positive integer. To obtain the maximum value;

[0074] Centered on each pixel, calculate the mean and standard deviation of brightness for each type of window to obtain the mean and standard deviation of window brightness.

[0075] The brightness difference value is obtained by normalizing the difference between the brightness of each pixel and the mean brightness of the window using the standard deviation of the window brightness.

[0076] In this embodiment, the formula for obtaining the brightness difference value is:

[0077] ,

[0078] in, For the first The first pixel The difference in brightness between the windows. It is a natural constant. For the first The brightness of each pixel For the first The first pixel The average brightness of the window type. For the first The first pixel The standard deviation of window brightness for each type of window. For the denominator parameter, and `|` represents a positive integer, and `|` represents the absolute value operation.

[0079] This invention obtains the brightness of a pixel by taking the maximum value of its R, G, and B channels and normalizing the result. This provides the basic brightness and darkness features of the pixel. Furthermore, it calculates the local brightness mean and standard deviation using a multi-scale m×m neighborhood window, and generates a brightness difference value by normalizing the difference between the pixel brightness and the window mean. This brightness difference value quantifies the degree of brightness aberration of a pixel relative to its neighborhood, making aberrant areas have a higher response in numerical space, while normal areas have a lower response because their brightness is close to the neighborhood average.

[0080] This invention, by calculating the mean / standard deviation of brightness under multiple windows, can simultaneously capture the differences between small windows (corresponding to tiny pinholes and localized missed spraying) and large windows (corresponding to large-area uneven spraying), solving the problem that traditional fixed-scale methods "can only adapt to defects of a single size" and achieving full coverage detection of defects of different sizes.

[0081] In this invention, the brightness difference in normal areas is compressed to near 0, weakening irrelevant fluctuations; the difference in real bright defects quickly approaches 1, highlighting defect features and effectively improving the distinguishability between bright defects and normal areas.

[0082] This invention extracts the darkness and brightness of each pixel, which is equivalent to obtaining a darkness distribution map and a brightness distribution map, describing the abnormal area from two dimensions. In contrast, the traditional single grayscale image is a weighted average of the RGB channels, which "neutralizes" the extreme brightness and darkness features of the pixels, causing the abnormal area to be covered up.

[0083] In this embodiment, the process of obtaining the darkness anomaly distribution map and the lightness anomaly distribution map includes:

[0084] For any pixel, if there exists a window whose darkness difference value is greater than a threshold, then the pixel is marked as a darkness outlier.

[0085] For any pixel, if there exists a window whose brightness difference value is greater than a threshold, then the pixel is marked as a brightness outlier.

[0086] The maximum value is selected from the darkness difference values ​​of multiple windows corresponding to the darkness anomaly point and used as the darkness difference representative value of the corresponding point. The darkness difference representative values ​​of other normal pixels are set to 0. The darkness difference representative values ​​are arranged according to the pixels to obtain the darkness anomaly distribution map.

[0087] The maximum value is selected from the brightness difference values ​​of multiple windows corresponding to the brightness anomaly points and used as the representative value of the brightness difference of the corresponding points. The representative values ​​of the brightness difference of other normal pixels are set to 0. The representative values ​​of brightness difference are arranged by pixel to obtain the brightness anomaly distribution map.

[0088] This invention uses the rule that "any difference value greater than a threshold is marked as an anomaly" to retain various dark / bright defects captured under different windows (such as tiny pinholes identified by small windows and large-area unevenness identified by large windows), avoiding the problem of missing defects of specific sizes by a single window and improving the defect detection rate.

[0089] This invention uses the "maximum difference value under multiple windows" as the final value for filtering outliers, which can highlight the "most significant degree of anomaly" of the point at all scales. For example, if a pixel has a difference value of 0.3 under a small window and 0.9 under a large window, it indicates that it has a defect under the large window. Taking 0.9 can more accurately reflect the severity of its anomaly and make the features of the defective area more prominent.

[0090] In this embodiment, the threshold is set to 0.5, but it can also be set to the average of the darkness difference or brightness difference values ​​of all pixels under the same window.

[0091] This invention sets the difference value of non-abnormal points to 0, which preserves the core features of abnormal regions and reduces the amount of image data for subsequent processing.

[0092] In this embodiment, the process of obtaining the darkness anomaly calibration map and the lightness anomaly calibration map includes:

[0093] The number of dark anomalies within a 3×3 neighborhood is counted, and the ratio of the number of dark anomalies to the number of pixels in the 3×3 neighborhood is taken as the region connectivity of the dark anomaly. Here, the number of pixels in the 3×3 neighborhood is equal to 9.

[0094] The darkness difference representative value of the same darkness anomaly point in the darkness anomaly distribution map is calibrated based on the regional connectivity of the darkness anomaly points to obtain the darkness anomaly calibration map.

[0095] ,

[0096] in, For the first Calibration value for the darkness difference of each darkness anomaly point. For the first The difference in darkness between each dark outlier point represents a value. For the first Regional connectivity of a dark anomaly point This represents the number of dark anomalies within a 3×3 neighborhood. It is a positive integer;

[0097] The number of brightness anomalies within a 3×3 neighborhood of a brightness anomaly is counted, and the ratio of the number of brightness anomalies to the number of pixels in the 3×3 neighborhood is taken as the region connectivity of that brightness anomaly.

[0098] The representative value of the brightness difference of the same brightness anomaly point in the brightness anomaly distribution map is calibrated based on the regional connectivity of the brightness anomaly points to obtain the brightness anomaly calibration map.

[0099] The calibration formula is:

[0100] ,

[0101] in, For the first The calibration value for the brightness difference of each brightness outlier point. For the first The brightness difference of each brightness outlier represents a value. For the first The regional connectivity of a brightness outlier point This represents the number of dark anomalies within a 3×3 neighborhood. It is a positive integer.

[0102] This invention achieves differentiated calibration by introducing the number N of neighborhood outliers: when ,use Amplify the difference value to enhance the salience of dense defects; when ,use The characteristics of sparse points are weakened to avoid over-amplification of isolated spurious anomalies. For points with low connectivity, the difference value is weakened (even approaching 0) after calibration, effectively filtering out irrelevant interference and making the detection results more accurate.

[0103] In this embodiment, the process of obtaining the brightness uniformity index distribution map includes:

[0104] At the same pixel location, if the corresponding pixel location in the darkness anomaly calibration map is a darkness anomaly point, or if the corresponding pixel location in the lightness anomaly calibration map is a lightness anomaly point, calculate the lightness / darkness consistency index: ,in, For the first The brightness uniformity index of each pixel For the first Calibration value for darkness difference of each pixel. For the first The brightness difference calibration value of each pixel, where e is the natural constant and || is the absolute value;

[0105] At the same pixel, the corresponding pixel position in the darkness anomaly calibration map is a normal pixel position, and the corresponding pixel position in the lightness anomaly calibration map is also a normal pixel position. Set the lightness and darkness consistency index at the corresponding pixel position to 0, and arrange each lightness and darkness consistency index according to the pixel position to obtain the lightness and darkness consistency index distribution map.

[0106] If, at the same pixel location, the corresponding pixel location in the darkness anomaly calibration map is a darkness anomaly point, or the corresponding pixel location in the lightness anomaly calibration map is a lightness anomaly point, it indicates that the coating area corresponding to that pixel point deviates from the normal lightness and darkness state, belonging to a potential area of ​​abnormal spraying quality. In this invention, when both the darkness difference calibration value and the lightness difference calibration value deviate in a consistent manner, the closer the lightness and darkness consistency index is to 1; when the two deviations are inconsistent, the lightness and darkness consistency index is smaller, indicating that there is inconsistency in both darkness and lightness dimensions. For spraying anomalies, both their darkness and lightness will be abnormal compared to the surrounding areas. Therefore, setting a lightness and darkness consistency index can improve the ability to accurately and effectively judge the true and effective areas of spraying anomalies.

[0107] like Figure 2 As shown, the guided enhancement quality assessment neural network includes: a first guided enhancement module, a second guided enhancement module, a first CNN network, a second CNN network, an adder, and a fully connected layer.

[0108] The output of the first guidance enhancement module is connected to the input of the first CNN network;

[0109] The output of the second guidance enhancement module is connected to the input of the second CNN network;

[0110] The input of the adder is connected to the output of the first CNN network and the output of the second CNN network, respectively, and its output is connected to the input of the fully connected layer.

[0111] The adder is used to add the output features of the first CNN network and the output features of the second CNN network element by element.

[0112] In this embodiment, the expression for the first boot enhancement module is:

[0113] ,

[0114] in, This is the darkness consistency enhancement feature output by the first guidance enhancement module. This is a calibration map for darkness anomalies. This is a distribution chart of the uniformity of light and dark indices. For element-wise multiplication, This is the first convolutional layer. This is the second convolutional layer;

[0115] The expression for the second boot enhancement module is:

[0116] ,

[0117] in, The brightness consistency enhancement feature output by the second guidance enhancement module. This is a brightness anomaly calibration chart. This is the third convolutional layer.

[0118] In this embodiment, the kernel size of the first, second, and third convolutional layers is 3×3.

[0119] In this embodiment, a score of 0.85–1.00 indicates excellent spraying quality, uniform coating color, and no significant fluctuations in shade; a score of 0.70–0.85 indicates good spraying quality, with small abnormal areas and weak impact on overall uniformity; a score of 0.50–0.70 indicates moderate deviation in spraying quality, with obvious abnormal reflections in local areas, and possible instances of spraying that is too thin, too thick, or partially uncovered; a score of 0.30–0.50 indicates poor spraying quality, with discontinuous coating texture, possibly accompanied by spraying breaks, particle accumulation, or uneven coverage; and a score of 0.00–0.30 indicates severe spraying abnormalities, potentially including major quality problems such as spray gun clogging, spray trajectory deviation, and large-scale missed spraying.

[0120] This invention utilizes the brightness consistency index distribution map as a guiding signal to adaptively enhance the features corresponding to the darkness anomaly calibration map and the brightness anomaly calibration map. Specifically, the first guided enhancement module element-wise multiplies the local structural features extracted from the darkness anomaly calibration map through convolution with the guiding features obtained by convolution of the consistency index, enabling the darkness anomaly region to obtain a higher response in the enhanced features; the second guided enhancement module performs the same targeted enhancement on the brightness anomaly calibration map. The first and second CNN networks set after the guided enhancement modules can perform deep feature extraction on the enhanced darkness and brightness features, then fuse the two types of features through an adder, and finally output the final coating quality score through a fully connected layer.

[0121] Existing methods rely on only a single window scale, resulting in poor adaptability to defects of different sizes. This invention extracts the dark / brightness difference value through multiple window scales while retaining the feature information of defects of different sizes. It can accurately identify both small-scale local defects and large-scale regional inhomogeneities, making defect detection more comprehensive and improving accuracy.

[0122] This invention calibrates anomaly maps by using regional connectivity, which can effectively distinguish between "real continuous defects" and "isolated noise points".

[0123] This invention introduces a brightness uniformity index to quantify the spatial matching degree between darkness anomalies and brightness anomalies, directly reflecting the quality problem of "uniform brightness" in the coating. Combined with the guided enhancement neural network for in-depth feature mining, it highlights the coupling relationship of the same location in different brightness directions, thereby significantly improving the discrimination significance of abnormal areas, suppressing background interference, and improving the accuracy of detection.

[0124] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for judging the quality of graphene ceramic coating spraying on a water turbine, characterized in that, Includes the following steps: Darkness and brightness were extracted from each pixel in the image of the graphene ceramic coating of the water turbine, and the differences in darkness and brightness were obtained under different window sizes: For each pixel's R, G, and B channel values, the minimum channel value is selected, and the ratio of the minimum channel value to 255 is taken as the darkness of the corresponding pixel. Centered on each pixel, the mean and standard deviation of darkness under each window are calculated to obtain the window darkness mean and window darkness standard deviation. Based on the difference between the darkness of each pixel and the window darkness mean, and normalized using the window darkness standard deviation, the darkness difference value under the corresponding window is obtained. For each pixel's R, G, and B channel values, select the maximum channel value and use the ratio of the maximum channel value to 255 as the brightness of the corresponding pixel. Centered on each pixel, calculate the mean and standard deviation of brightness for each type of window to obtain the mean and standard deviation of window brightness; based on the difference between the brightness of each pixel and the mean brightness of the window, normalize using the standard deviation of window brightness to obtain the brightness difference value. Based on multiple darkness difference values ​​and multiple lightness difference values ​​of the same pixel, darkness anomalies and lightness anomalies are marked, and the maximum value of multiple darkness difference values ​​and lightness difference values ​​of the anomalies is selected to obtain darkness anomaly distribution map and lightness anomaly distribution map. The number of dark anomalies within a 3×3 neighborhood of a dark anomaly is counted, and the ratio of the number of dark anomalies to the number of pixels in the 3×3 neighborhood is taken as the region connectivity of that dark anomaly. Based on the regional connectivity of the dark anomalies, the representative values ​​of the dark difference at the same dark anomaly point in the dark anomaly distribution map are calibrated to obtain the dark anomaly calibration map: ,in, For the first Calibration value for the darkness difference of each darkness anomaly point. For the first The difference in darkness between each dark outlier point represents a value. For the first Regional connectivity of a dark anomaly point This represents the number of dark anomalies within a 3×3 neighborhood. It is a positive integer; The number of brightness anomalies within a 3×3 neighborhood of a brightness anomaly is counted, and the ratio of the number of brightness anomalies to the number of pixels in the 3×3 neighborhood is taken as the region connectivity of that brightness anomaly. Based on the regional connectivity of the lightness anomalies, the representative value of the lightness difference at the same lightness anomaly point in the lightness anomaly distribution map is calibrated to obtain the lightness anomaly calibration map: ,in, For the first The calibration value for the brightness difference of each brightness outlier point For the first The brightness difference of each brightness outlier represents a value. For the first Regional connectivity of a brightness outlier point This represents the number of dark anomalies within a 3×3 neighborhood. It is a positive integer; Based on the difference between the shadow anomaly calibration map and the lightness anomaly calibration map, a lightness uniformity index is generated, resulting in a lightness uniformity index distribution map: ,in, For the first The brightness uniformity index of each pixel For the first Calibration value for darkness difference of each pixel. For the first The brightness difference calibration value of each pixel, where e is the natural constant and || is the absolute value; A guided enhanced quality assessment neural network was used to process the dark anomaly calibration map and the light anomaly calibration map, and feature enhancement was performed based on the light and dark consistency index distribution map to obtain the coating quality score.

2. The method for judging the coating quality of graphene ceramic gold coating for water turbines according to claim 1, characterized in that, The formula for obtaining the difference in darkness is: ， in, For the first The first pixel The difference in darkness between the windows. It is a natural constant. For the first The darkness of each pixel For the first The first pixel The average window darkness of each type of window. For the first The first pixel The standard deviation of window darkness for this type of window. For the denominator parameter, and `|` represents a positive integer, and `|` represents the absolute value operation.

3. The method for judging the coating quality of graphene ceramic gold coating for water turbines according to claim 1, characterized in that, The formula for obtaining the brightness difference value is: ， in, For the first The first pixel The difference in brightness between the windows. It is a natural constant. For the first The brightness of each pixel For the first The first pixel The average brightness of the window type. For the first The first pixel The standard deviation of window brightness for each type of window. For the denominator parameter, and `|` represents a positive integer, and `|` represents the absolute value operation.

4. The method for judging the coating quality of graphene ceramic gold coating for water turbines according to claim 1, characterized in that, The process of obtaining the darkness anomaly distribution map and the lightness anomaly distribution map includes: For any pixel, if there exists a window whose darkness difference value is greater than a threshold, then the pixel is marked as a darkness outlier. For any pixel, if there exists a window whose brightness difference value is greater than a threshold, then the pixel is marked as a brightness outlier. The maximum value is selected from the darkness difference values ​​of multiple windows corresponding to the darkness anomaly points and used as the representative value of the darkness difference of the corresponding points. The other normal pixels are set to 0 to obtain the darkness anomaly distribution map. The maximum value is selected from the brightness difference values ​​of multiple windows corresponding to the brightness anomaly points and used as the representative value of the brightness difference of the corresponding points. The other normal pixels are set to 0 to obtain the brightness anomaly distribution map.

5. The method for judging the coating quality of graphene ceramic gold coating for water turbines according to claim 1, characterized in that, The process of obtaining the uniformity of light and dark index distribution map includes: At the same pixel, the corresponding pixel position in the darkness anomaly calibration map is a darkness anomaly point, or the corresponding pixel position in the lightness anomaly calibration map is a lightness anomaly point; calculate the lightness and darkness consistency index. At the same pixel, the corresponding pixel position in the darkness anomaly calibration map is a normal pixel position, and the corresponding pixel position in the brightness anomaly calibration map is also a normal pixel position. Set the brightness uniformity index of the corresponding pixel position to 0 to obtain the brightness uniformity index distribution map.

6. The method for judging the coating quality of graphene ceramic gold coating for water turbines according to claim 1, characterized in that, The guided enhancement quality assessment neural network includes: a first guided enhancement module, a second guided enhancement module, a first CNN network, a second CNN network, an adder, and a fully connected layer.

7. The method for judging the coating quality of graphene ceramic gold coating for water turbines according to claim 6, characterized in that, The expression for the first boot enhancement module is: ， in, This is the darkness consistency enhancement feature output by the first guidance enhancement module. This is a calibration map for darkness anomalies. This is a distribution chart of the uniformity of light and dark indices. For element-wise multiplication, This is the first convolutional layer. This is the second convolutional layer; The expression for the second boot enhancement module is: ， in, The brightness consistency enhancement feature output by the second guidance enhancement module. This is a brightness anomaly calibration chart. This is the third convolutional layer.

Images